Schematic representation of the seven dimensions of hardware requirements for truly brain-like computing. Reproduced with permission from: Terasa et al., Materials Today (2023), http://doi.org/10.1016/j.mattod.2023.07.019.
Schematic of the exploration space for the development of bio-inspired computing architectures. (© Martin Ziegler and Hermann Kohlstedt, Kiel University.)Despite remarkable progress in digital computing, biological systems remain unmatched in cognitive tasks such as pattern recognition at low power consumption. Inspired by the information processing capabilities of neural assemblies, researchers from Kiel University have identified materials systems and the essential mechanisms for the development of novel brain-like computing [Terasa et al., Materials Today (2023), http://doi.org/10.1016/j.mattod.2023.07.019 ].
“Brain-like computing differs considerably from digital computing,” explains Hermann Kohlstedt, who led the study with Rainer Adelung and Alexander Vahl. “Information processing in biological systems – including our brain – is characterized by parallel processing. This is in stark contrast to digital computing where processing is serial.”
Where digital computing is highly energy intensive the human brain, by contrast, consumes just 25 W during complex visual or sound recognition tasks. Mimicking the basic principles of biological information processing in materials systems could be translated into a new kind of electronics for future prosthetics, robotics, or autonomous vehicles.
“The goal of our experimental approach is to explore novel pathways of information processing inspired by biological systems,” adds Kohlstedt. “The platform of future electronics could differ considerably from today’s silicon technology, however.”
Firstly, biological processing systems are three dimensional and exploit a complex network of synapses while digital processors are mainly two dimensional. The connections in biological systems change and rearrange over time. This ‘distributed plasticity’ is one of the key requirements for brain-like computing. Different approaches have the potential to achieve dynamically reconfigurable connections including nano-granular networks of nanoparticles, liquid-solid composites, directed electrochemical growth of metal filaments, and relaxation-type oscillators in a liquid matrix.
The concept of these in materia systems is to ‘let the material compute’, with self-organizing nanowires or nanoparticles with electrical in- and outputs performing dynamic operations. Such systems must balance plasticity with stability, operating near criticality. As neural networks develop, the self-ordered arrangement of the system becomes important with connections bridging larger gaps within the system. Maintaining both long- and short-range connections – or hierarchy and modularity – is crucial for better information processing. Finally, brain-like processing systems must be robust, truly 3D, and able to orchestrate the coupling of ensembles of neuron-like structures.
“On the one hand, we are far from being competitive with current artificial intelligence,” points out Kohlstedt, “nonetheless, applying only a few principles such as criticality and plasticity enable very interesting sensor systems.”